Open Bandit Pipeline: Logged Bandit Feedback for Off-Policy Evaluation

Open Bandit Pipeline provides logged bandit feedback data and a Python framework for off-policy evaluation, developed by st-tech and updated in June 2024. The data facilitates benchmarking bandit algorithms using real-world interaction logs from the ZOZOTOWN e-commerce platform, though specific record counts are not provided in the metadata.

Use Cases

• Benchmarking off-policy evaluation (OPE) estimators using logged reward signals
• Training contextual bandit algorithms on historical action-reward pairs
• Simulating bandit environments to compare policy performance against logged data

Strengths

• Apache-2.0 licensed for open research
• Includes real-world logged bandit feedback from an e-commerce environment
• Designed specifically for off-policy evaluation (OPE) benchmarking

Limitations

• Specific column names and row counts are not provided in the metadata
• Domain-specific to e-commerce recommendation systems

Provenance

Source

st-tech

Collection Method

Logged interactions from a production recommendation system

Freshness

Last updated June 2024.